A product quality index prediction method and system based on AM-GRU-BPNN

Abstract

The present invention discloses a product quality index prediction method and system based on AM‑GRU‑BPNN. The method includes: collecting multiple groups of working condition data from multiple processes of a process manufacturing production line as a sample data set; preprocessing each group of sample data in the sample data set; dividing into training set data and test set data; constructing an AM‑GRU‑BPNN prediction model, training the AM‑GRU‑BPNN prediction model, and obtaining a trained AM‑GRU‑BPNN prediction model; and testing the test set data using the trained AM‑GRU‑BPNN prediction model. The present invention specifically solves the problem of difficult product quality prediction in the process manufacturing industry, can predict defects in advance and make up for them in time, thereby improving the prediction accuracy of quality indicators, reducing enterprise costs, and providing methods and approaches for realizing accurate prediction and optimized regulation of process production processes.

Description

Technical Field

[0001] The invention relates to a product quality index prediction method and system based on AM-GRU-BPNN, belonging to the technical field of time series prediction. Background Art

[0002] Unlike traditional discrete manufacturing, the equipment characteristics, process formulas, raw material properties, and control parameters of each process in a process-based industrial production line are closely related to production performance indicators. Therefore, how to interact and integrate the real-time data of production factors such as equipment, materials, and products, realize on-demand response, rapid iteration, and dynamic optimization of resource allocation and operation within the system, and improve the predictability and linkage of production activity plans are the core issues in reshaping the intelligent production model of process industries. Therefore, it is necessary to start from solving the bottleneck problems of self-perception, self-calculation, and self-prediction of all process elements in the multi-process coupling production process of process manufacturing in an intelligent manufacturing environment, so as to predict product quality problems in advance before the finished product is produced, reduce enterprise costs, and provide methods and ways to achieve accurate prediction and optimal control of process production processes. Summary of the invention

[0003] The present invention provides a product quality index prediction method and system based on AM-GRU-BPNN, which are used to construct an AM-GRU-BPNN model to realize the prediction of product quality indicators in a process manufacturing production line.

[0004] The technical solution of the present invention is: a product quality index prediction method based on AM-GRU-BPNN, comprising:

[0005] Collect multiple groups of working condition data from multiple processes of a process manufacturing production line as sample data sets; wherein the working condition data include process parameter data and quality index data;

[0006] Preprocess each set of sample data in the sample data set to obtain non-redundant data;

[0007] Divide the non-redundant data into training set data and test set data according to the proportion and time order;

[0008] Constructing an AM-GRU-BPNN prediction model, training the AM-GRU-BPNN prediction model, and obtaining a trained AM-GRU-BPNN prediction model; wherein the AM-GRU-BPNN prediction model includes a temporal attention mechanism AM, a gated recurrent unit network prediction model GRU, and a BP neural network prediction model;

[0009] The trained AM-GRU-BPNN prediction model is used to test the test set data to obtain the prediction results.

[0010] The preprocessing includes: firstly filling the missing values ​​in the sample data in the sample data set with the mean, and then eliminating the sample data using the 3sigma criterion; normalizing the sample data set judged by the 3sigma criterion; quantitatively analyzing the influence of process parameters on quality indicators based on information entropy, and using the information entropy threshold to realize feature screening of normalized data to obtain non-redundant data.

[0011] The AM-GRU-BPNN prediction model is constructed, and the AM-GRU-BPNN prediction model is trained to obtain the trained AM-GRU-BPNN prediction model, including:

[0012] Combine the training set data at time t with the hidden state h at the previous time t-1 Combined with the input of the temporal attention mechanism AM, the weight of the working condition data corresponding to the working condition data feature at time t is obtained by calculation Multiply the working condition data weight corresponding to the working condition data feature at time t by the corresponding working condition data feature to obtain the associated feature x of the working condition data feature t ';

[0013] The gated recurrent unit network prediction model GRU includes an input layer, a hidden layer, and an output layer. The output x of the temporal attention mechanism AM is t ' is used as the input of the gated recurrent unit network prediction model GRU, and the reconstructed working condition data is obtained by calculation. The reconstructed working condition data is used as the input of the hidden layer. By updating the GRU hidden layer state at time t, the processed sliding data X' is obtained as the output of GRU, and T groups of sliding data are obtained;

[0014] The sliding data X' output by the gated recurrent unit network prediction model GRU is used as the input of the BP neural network prediction model, and the input matrix c of the BP neural network prediction model at time t is used as the input matrix c of the BP neural network prediction model at time t. 1 ,c 2 ,...,c T , the input matrix and the introduced weight matrix W 1 ,W 2 ,...,W T Combined to get the output correlation degree Multiply the output correlation degree with the BP neural network prediction model input matrix corresponding to time t to obtain the process weight θ f ; Normalize the process weights of f processes to obtain the normalized process weights Normalize the process weights It is regarded as a neuron, connected to the fully connected layer network and performs back propagation calculation, and the parameters in the preset neural network model are updated to obtain a trained AM-GRU-BPNN model.

[0015] The number of neurons in the hidden layer of the gated recurrent unit network prediction model GRU is 64, 20, and 50.

[0016] The method uses the trained AM-GRU-BPNN prediction model to test the test set data to obtain the prediction results. Specifically, the trained AM-GRU-BPNN prediction model is tested on the test set data with other models to obtain the prediction results, and the evaluation indicators of the prediction results are compared and analyzed.

[0017] The other models are one or more of an AM-GRU model, a GRU model, a LSTM model, and a BPNN model.

[0018] According to another aspect of the present invention, there is also provided a prediction system for multi-process process parameters, comprising:

[0019] A collection module is used to collect multiple groups of working condition data from multiple processes of a process manufacturing production line as sample data sets; wherein the working condition data includes process parameter data and quality index data;

[0020] The first acquisition module is used to preprocess each set of sample data in the sample data set to obtain non-redundant data;

[0021] A partitioning module is used to divide the non-redundant data into training set data and test set data according to the proportion and time order;

[0022] A training module is used to construct an AM-GRU-BPNN prediction model, train the AM-GRU-BPNN prediction model, and obtain a trained AM-GRU-BPNN prediction model; wherein the AM-GRU-BPNN prediction model includes a temporal attention mechanism AM, a gated recurrent unit network prediction model GRU, and a BP neural network prediction model;

[0023] The second acquisition module uses the trained AM-GRU-BPNN prediction model to test the test set data and obtain the prediction results.

[0024] According to another aspect of the present invention, a terminal device is also provided, including a memory, a processor, and a program stored in the memory and executable by the processor, wherein when the processor executes the program, the product quality index prediction method based on AM-GRU-BPNN described in any one of the above is implemented.

[0025] According to another aspect of the present invention, a computer-readable storage medium is also provided, which includes a stored program, wherein when the program is running, the device where the computer-readable storage medium is located is controlled to execute any one of the above-mentioned product quality indicator prediction methods based on AM-GRU-BPNN.

[0026] The beneficial effects of the present invention are as follows: the present invention specifically solves the problem of difficult product quality prediction in the process manufacturing industry, that is, product quality problems can be predicted in advance before the finished product is produced, defects can be predicted in advance and remedied in time, thereby improving the prediction accuracy of quality indicators, reducing enterprise costs, and providing methods and approaches for realizing accurate prediction and optimized regulation of process production processes.

[0027] Specifically: the present invention preprocesses the sample data, including filling and removing the sample data, normalizing and feature screening. By filling and removing the sample data, the sample data can be made more stable, laying the foundation for the subsequent construction of a high-precision prediction model; normalizing the data to between 0 and 1, saving computing resources and accelerating the convergence speed of the neural network; quantitatively analyzing the influence of process parameters on quality indicators through information entropy, realizing feature screening of high-dimensional process parameters, and solving the problem of feature extraction of process parameter features with high-dimensional time series in the production process; the present invention introduces a time series attention mechanism for process parameter features, and realizes autonomous processing of process parameter time series data by allocating attention weights to process parameter feature vectors, solving the short-term memory problem of recurrent neural networks; increasing the depth of the model by adding GRU network units helps to improve the prediction ability of the model; and further cooperating with the BPNN model to obtain process weights strengthens the expression of the quality indicator output by the time series information of the process parameters, and improves the stability of the quality prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic diagram of the process of the present invention;

[0029] Figure 2 This is the internal unit structure diagram of the temporal attention mechanism AM of the present invention;

[0030] Figure 3 This is a diagram of the internal unit structure of the gated recurrent unit network GRU of the present invention;

[0031] Figure 4 It is the internal unit structure diagram of BPNN of the present invention;

[0032] Figure 5 This is a schematic diagram comparing the predicted value curve and the actual value curve of the quality index predicted on the test set using the AM-GRU-BPNN model in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] The invention is further described below in conjunction with the accompanying drawings and embodiments, but the content of the invention is not limited to the scope of the embodiments.

[0034] Example 1: Figure 1-5 As shown, a product quality index prediction method based on AM-GRU-BPNN includes:

[0035] Collect multiple groups of working condition data from multiple processes of a process manufacturing production line as sample data sets; wherein the working condition data include process parameter data and quality index data;

[0036] Preprocess each set of sample data in the sample data set to obtain non-redundant data; by preprocessing the sample data, the sample data can be made more stable; and the use of stable sample data lays the foundation for the subsequent construction of a high-precision prediction model.

[0037] According to the proportion and time order, the non-redundant data is divided into training set data Train = (X train , Y train ) and test set data Test=(X test , Y test ); where the training set data Train = (X train , Y train ), for each set of training data, {x (1) ,x (2) ,…,x (n)}, that is, n-1-dimensional process parameter data and 1-dimensional quality index data, that is, a set of training set data has n dimensions;

[0038] Constructing an AM-GRU-BPNN prediction model, training the AM-GRU-BPNN prediction model, and obtaining a trained AM-GRU-BPNN prediction model; wherein the AM-GRU-BPNN prediction model includes a temporal attention mechanism AM, a gated recurrent unit network prediction model GRU, and a BP neural network prediction model;

[0039] The trained AM-GRU-BPNN prediction model is used to test the test set data to obtain the prediction results.

[0040] Furthermore, the preprocessing can be set to include: first filling the missing values ​​in the sample data in the sample data set with the mean, and then eliminating the sample data using the 3sigma criterion; normalizing the sample data set judged by the 3sigma criterion; quantitatively analyzing the impact of process parameters on quality indicators based on information entropy, and using the information entropy threshold to achieve feature screening of normalized data to obtain non-redundant data.

[0041] Specifically, the normalization process refers to mapping the sample data to between [0, 1], and the normalization formula is as follows:

[0042]

[0043] Where: x i * indicates normalized data; x max 、x min Respectively represent the maximum and minimum values ​​of the sample data before normalization in the sample data set; x i Represents the original sample data;

[0044] The information entropy calculation formula is as follows:

[0045]

[0046] Where: I(X; Y) represents the information entropy between process parameter X and quality index Y; p(x) and p(y) represent the marginal density functions of process parameter X and quality index Y respectively; p(x, y) represents the joint probability density function of process parameter X and quality index Y; the larger the entropy value, the stronger the correlation between the two.

[0047] Furthermore, if Figure 2-Figure 4 As shown, the AM-GRU-BPNN prediction model is constructed, and the AM-GRU-BPNN prediction model is trained to obtain the trained AM-GRU-BPNN prediction model, including:

[0048] Combine the training set data at time t with the hidden state h at the previous time t-1 Combined with the input of the temporal attention mechanism AM, the weight of the working condition data corresponding to the working condition data feature at time t is obtained by calculation Multiply the working condition data weight corresponding to the working condition data feature at time t by the corresponding working condition data feature to obtain the associated feature x of the working condition data feature t ', AM output network unit is denoted as f GRU1 ; Among them, by continuously correlating the features x t 'Iteration can ensure that the hidden state h at each moment t Both contain the associated feature x t ';

[0049] The gated recurrent unit network prediction model GRU includes an input layer, a hidden layer, and an output layer. The output x of the temporal attention mechanism AM is t 'As the input of the gated recurrent unit network prediction model GRU, for the network unit f GRU1The reconstructed working condition data is calculated and used as the input of the hidden layer. By updating the GRU hidden layer state at time t, the processed sliding data X' is obtained as the output of GRU, and T groups of sliding data are obtained. The GRU output network unit is recorded as f GRU2 ; The number of neurons in the input layer is 26, and the number of neurons in the hidden layer is 64, 20, and 50; Experiments have shown that setting the number of neurons in the hidden layer to 64, 20, and 50 can make the calculation faster and prevent overfitting.

[0050] The sliding data X' output by the gated recurrent unit network prediction model GRU is used as the input of the BP neural network prediction model, and the input matrix c of the BP neural network prediction model at time t is used as the input matrix c of the BP neural network prediction model at time t. 1 ,c 2 ,...,c T , c T =f GRU2 (c T-1 ,X'), the input matrix and the introduced weight matrix W 1 ,W 2 ,...,W T Combined to get the output correlation degree Multiply the output correlation degree with the BP neural network prediction model input matrix corresponding to time t to obtain the process weight θ f ; Normalize the process weights of f processes to obtain the normalized process weights Normalize the process weights It is regarded as a neuron, connected to the fully connected layer network and performs back propagation calculation, and the parameters in the preset neural network model are updated to obtain a trained AM-GRU-BPNN model.

[0051] Specifically, the AM-GRU-BPNN prediction model of the present invention first introduces AM so that the model can extract data features at a deeper level and improve the prediction accuracy of the model; it is placed after GRU according to the characteristics of BPNN: one is based on the characteristics of multiple processes, it is more in line with the expression of actual process; the other is to highlight the advantages of BPNN connecting the fully connected layer, making the calculation of the entire model more efficient.

[0052] Furthermore, the working condition data weight calculation formula is as follows:

[0053]

[0054] Where: W represents the weight of the working condition data corresponding to the nth working condition data feature at time t; t (n) represents the weight matrix corresponding to the nth operating condition data feature; h t-1represents the hidden state of the previous moment at the working condition data feature time t; B t represents the deviation term corresponding to the working condition data feature at time t; Selu represents the activation function;

[0055] The calculation formula of the association feature is as follows:

[0056]

[0057] Where: x' t Represents the correlation features of different working condition data features at time t; Represents the kth operating condition data characteristics at time t; Represents the operating condition data weight corresponding to the kth operating condition data feature at time t.

[0058] The calculation formula for continuously iterating the associated features is as follows:

[0059] h t =f GRU1 (h t-1 ,x' t )

[0060] Where: h t represents the hidden state of AM at time t; x' t Represents the associated features of the working condition data features; f GRU1 Represents the input side network unit of the gated recurrent unit network prediction model GRU;

[0061] The GRU training calculation formula is as follows:

[0062] r t =Sigmoid(W xz x t +W hz h' t-1 +b r )

[0063] z t =Sigmoid(W xz x t +W hz h' t-1 +b z )

[0064]

[0065]

[0066] X'=σ(W ho h' t +b y )

[0067] Where: r t Represents the reset gate; Z t represents the update gate; represents the hidden gate; σ represents the activation function Sigmoid; W xz Represents the GRU network input layer and Z t Calculate weight; W xh Represents the GRU network input layer and r at time t t Calculate weight; W hz Represents the hidden layer at time t-1 and Z t Calculate weight; W hr Represents the hidden layer at time t-1 and r t Calculate weight; W ho represents the weights calculated between the hidden layer and the GRU network output layer at time t, b h represents the calculation bias of the hidden layer of the GRU network at time t, b y Represents the calculation bias of the GRU network output layer at time t.

[0068] The back propagation calculation formula is as follows:

[0069]

[0070] Where: Represents the predicted value of the quality index of the training set data; W y Indicates the calculation weight; b' y Indicates the BPNN network calculation bias;

[0071] Furthermore, the trained AM-GRU-BPNN prediction model can be set to test the test set data to obtain prediction results. Specifically, the trained AM-GRU-BPNN prediction model is tested on the test set data separately with other models to obtain prediction results, and the evaluation indicators of the prediction results are compared and analyzed.

[0072] Furthermore, the other models may be set to be one or more of an AM-GRU model, a GRU model, a LSTM model, and a BPNN model.

[0073] Specifically, the prediction results are calculated using mean absolute error (MAE), mean square error (MSE), and goodness of fit (R). 2 The calculation formula is as follows:

[0074]

[0075]

[0076]

[0077] Where: m represents the sample size of the test data set; and j represents the actual value and predicted value of the quality index at the jth moment; Represents the average value of the predicted value sample and the actual value, where j = 1, 2, ..., m; the smaller the MAE and MSE values, the higher the R 2 The closer the value is to 1, the greater the goodness of fit and the more accurate the model prediction results.

[0078] Illustratively, an optional implementation manner of the present invention is described in detail below, but the content of the present invention is not limited to the scope.

[0079] This embodiment aims at the high-dimensional, time-series and coupling characteristics of multi-process data in process-type manufacturing, and the existing machine learning prediction methods have the insufficiency of selecting key parameters based on experience, and proposes an intelligent multi-process process parameter association prediction method based on AM-GRU-BPNN. First, the mutual information method is used to screen the polymorphic heterogeneous production data as input, and it is constructed into a time series input to the attention mechanism-gated recurrent unit network-BP neural network prediction model for training and calculation, and finally the prediction of the moisture content quality index of the discharge is completed.

[0080] The sample data set uses the silk production line of a process manufacturing enterprise from September 7, 2021 to September 14, 2021, with a data collection cycle of 15s, and a total of 2432 groups of sample data. Each group of sample data consists of 72-dimensional process parameter data and 1-dimensional quality index data; each group of sample data in the sample data set is preprocessed, 432 groups of sample data are eliminated, and then the influence of process parameters on quality indicators is quantitatively analyzed based on information entropy, and the information entropy threshold is determined to be 0.25. Process parameters with information entropy greater than 0.25 are selected, thereby obtaining 25-dimensional process parameter features, realizing feature screening of high-dimensional process parameters, and finally obtaining 2000 groups of sample data consisting of 25-dimensional process parameters and 1-dimensional quality index data. The data is distributed in a ratio of 8:2, and the last 400 groups of the data set are used as a 20% test set. The above sample data are used for prediction under different models to verify the intelligent multi-process process parameter association prediction method based on AM-GRU-BPNN proposed in this invention.

[0081] The prediction model constructed by the attention mechanism-gated recurrent unit network-BP neural network prediction method is obtained by training the training set data. The settings of the model parameters are shown in Table 1.

[0082] Table 1 Parameters of the attention mechanism-gated recurrent unit network-BP neural network model

[0083] parameter Value (layer) Dimensions 26 Temporal Attention Mechanism Neurons 26 GRU input neurons 26 GRU hidden layer neurons 64、20、50 BPNN input layer neurons 6 BPNN output layer neurons 1

[0084] Combined with the evaluation index, the AM-GRU-BPNN model proposed in the present invention is used to predict the moisture content of the discharge material of the quality index with the AM-GRU model, the GRU model, the LSTM model and the BPNN model. The evaluation index is shown in Table 2. The trained AM-GRU-BPNN model is respectively compared with the AM-GRU model, the GRU model, the LSTM model and the BPNN model in the test set data Test = (X test , Y test ) and predict the quality index values ​​for the next 400 moments, and then calculate the MAE, MSE and R of the test results 2 The three evaluation indicators of each model are compared and analyzed. The analysis shows that compared with the AM-GRU, GRU, LSTM, and BPNN prediction models, the AM-GRU-BPNN method proposed in the present invention has reduced MAE indicators by 1.11%, 1.62%, 2.6%, and 0.64%, respectively, and MSE indicators by 1.16%, 1.51%, 0.18%, and 0.94%, respectively. This shows that the AM-GRU-BPNN model has better prediction effects in dealing with time series problems, reflects the necessity of introducing the attention mechanism, and effectively improves the model's ability to extract complex features of production process parameters. Finally, the effects of the five prediction models are quantitatively analyzed, and the determination coefficient R 2 The calculation results show that the R of the AM-GRU-BPNN prediction model proposed in this invention is 2 The value is 0.9717, which is closest to 1, which means that the model has the best prediction effect.

[0085] Table 2 Comparison of different prediction models

[0086] Prediction Model MAE / % MSE / % <![CDATA[R 2 ]]> AM-GRU-BPNN 1.43 0.07 0.9717 AM-GRU 2.54 1.23 0.9313 GRU 3.05 1.58 0.9047 LSTM 4.03 0.25 0.9158 BPNN 2.07 1.01 0.9425

[0087] The change curve of the moisture content quality index of the discharge material predicted by the attention mechanism-gated recurrent unit network-BP neural network (AM-GRU-BPNN) model proposed in the present invention and the actual value is as follows: Figure 5 As shown in the figure, it can be seen that the AM-GRU-BPNN model proposed in the present invention has the highest prediction accuracy, has a better fitting ability for the actual discharge moisture content quality index, and is basically consistent with the change trend of the actual discharge moisture content.

[0088] Through the above analysis, it can be known that in order to verify the effectiveness of the method of the present invention, this embodiment adopts the AM-GRU-BPNN model of the method of the present invention to predict quality indicators on the test set of the same data with the AM-GRU model, GRU model, LSTM model, and BPNN model respectively. The results show that compared with the LSTM model, the GRU model has a simpler structure, which improves the computing power of the model and reduces the computing cost; compared with the BPNN model, the BPNN model has good mapping and generalization capabilities, which helps to improve the prediction ability of the model; compared with the GRU model, the AM-GRU model has a temporal attention mechanism AM that realizes autonomous processing of process parameter data, fully extracts the characteristics of working condition data, and solves the problem of insufficient prediction accuracy caused by insufficient extraction of process parameter characteristics in the prior art. Therefore, the AM-GRU-BPNN model can maximize the advantages of the temporal attention mechanism AM, GRU model and BPNN model, effectively handle the complex relationship between process parameters and quality indicators, optimize the model parameters, realize multi-process production quality control, improve the prediction accuracy of production line process quality parameters, and make the prediction results of quality indicators have high accuracy. By using the method of the present invention, variable characteristics can be effectively mined, and the bottleneck problems of self-perception, self-calculation and self-prediction of all process elements in the multi-process coupled production process of the process manufacturing process under the intelligent manufacturing environment can be solved, thereby improving the prediction accuracy of quality indicators, reducing enterprise costs, and providing methods and approaches for realizing accurate prediction and optimal control of the process production process.

[0089] Embodiment 2:

[0090] According to another aspect of an embodiment of the present invention, a prediction system for multi-process process parameters is provided, including:

[0091] A collection module is used to collect multiple groups of working condition data from multiple processes of a process manufacturing production line as sample data sets; wherein the working condition data includes process parameter data and quality index data;

[0092] The first acquisition module is used to preprocess each set of sample data in the sample data set to obtain non-redundant data;

[0093] A partitioning module is used to divide the non-redundant data into training set data and test set data according to the proportion and time order;

[0094] A training module is used to construct an AM-GRU-BPNN prediction model, train the AM-GRU-BPNN prediction model, and obtain a trained AM-GRU-BPNN prediction model; wherein the AM-GRU-BPNN prediction model includes a temporal attention mechanism AM, a gated recurrent unit network prediction model GRU, and a BP neural network prediction model;

[0095] The second acquisition module uses the trained AM-GRU-BPNN prediction model to test the test set data and obtain the prediction results.

[0096] It should be noted that the above-mentioned modules can be implemented by software or hardware. For example, for the latter, it can be implemented in the following ways: the above-mentioned modules can be located in the same processor; and / or the above-mentioned modules are located in different processors in any combination.

[0097] Embodiment 3:

[0098] According to another aspect of an embodiment of the present invention, a terminal device is also provided, including a memory, a processor, and a program stored in the memory and executable by the processor, wherein when the processor executes the program, the product quality index prediction method based on AM-GRU-BPNN described in any one of the above is implemented.

[0099] Embodiment 4:

[0100] According to another aspect of an embodiment of the present invention, a computer-readable storage medium is also provided, wherein the computer-readable storage medium includes a stored program, wherein when the program is running, the device where the computer-readable storage medium is located is controlled to execute any one of the above-mentioned product quality indicator prediction methods based on AM-GRU-BPNN.

[0101] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0102] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0103] The specific implementation modes of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above implementation modes, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A product quality index prediction method based on AM-GRU-BPNN, characterized by: include: Collect multiple groups of working condition data from multiple processes of a process manufacturing production line as sample data sets; wherein the working condition data include process parameter data and quality index data; Preprocess each set of sample data in the sample data set to obtain non-redundant data; Divide the non-redundant data into training set data and test set data according to the proportion and time order; Constructing an AM-GRU-BPNN prediction model, training the AM-GRU-BPNN prediction model, and obtaining a trained AM-GRU-BPNN prediction model; wherein the AM-GRU-BPNN prediction model includes a temporal attention mechanism AM, a gated recurrent unit network prediction model GRU, and a BP neural network prediction model; Use the trained AM-GRU-BPNN prediction model to test the test set data and obtain the prediction results; The preprocessing includes: firstly filling the missing values ​​in the sample data in the sample data set with the mean value, and then eliminating the sample data by using the 3sigma criterion; normalizing the sample data set judged by the 3sigma criterion; quantitatively analyzing the influence of the process parameters on the quality index based on the information entropy, and using the information entropy threshold to realize the feature screening of the normalized data to obtain non-redundant data; The AM-GRU-BPNN prediction model is constructed, and the AM-GRU-BPNN prediction model is trained to obtain the trained AM-GRU-BPNN prediction model, including: Combine the training set data at time t with the hidden state h at the previous time t-1 Combined with the input of the temporal attention mechanism AM, the weight of the working condition data corresponding to the working condition data feature at time t is obtained by calculation Multiply the working condition data weight corresponding to the working condition data feature at time t by the corresponding working condition data feature to obtain the associated feature x of the working condition data feature t '; The gated recurrent unit network prediction model GRU includes an input layer, a hidden layer, and an output layer. The output x t ' is used as the input of the gated recurrent unit network prediction model GRU, and the reconstructed working condition data is obtained by calculation. The reconstructed working condition data is used as the input of the hidden layer. By updating the GRU hidden layer state at time t, the processed sliding data X' is obtained as the output of GRU, and T groups of sliding data are obtained; The sliding data X' output by the gated recurrent unit network prediction model GRU is used as the input of the BP neural network prediction model, and the input matrix c1, c2, ..., c T , the input matrix and the introduced weight matrix W1,W2,...,W T Combined to get the output correlation degree Multiply the output correlation degree with the BP neural network prediction model input matrix corresponding to time t to obtain the process weight θ f ; Normalize the process weights of f processes to obtain the normalized process weights Normalize the process weights It is regarded as a neuron, connected to the fully connected layer network and performs back propagation calculations to update the parameters in the preset neural network model to obtain the trained AM-GRU-BPNN model.

2. The product quality index prediction method based on AM-GRU-BPNN according to claim 1 is characterized in that: The number of neurons in the hidden layer of the gated recurrent unit network prediction model GRU is 64, 20, and 50.

3. The product quality index prediction method based on AM-GRU-BPNN according to claim 1 is characterized in that: The method uses the trained AM-GRU-BPNN prediction model to test the test set data to obtain the prediction results. Specifically, the trained AM-GRU-BPNN prediction model is tested on the test set data with other models to obtain the prediction results, and the evaluation indicators of the prediction results are compared and analyzed.

4. The product quality index prediction method based on AM-GRU-BPNN according to claim 3 is characterized in that: The other models are one or more of an AM-GRU model, a GRU model, a LSTM model, and a BPNN model.

5. A prediction system for multi-process parameters for implementing the product quality index prediction method based on AM-GRU-BPNN as claimed in claim 1, characterized in that: include: A collection module is used to collect multiple groups of working condition data from multiple processes of a process manufacturing production line as sample data sets; wherein the working condition data includes process parameter data and quality index data; The first acquisition module is used to preprocess each set of sample data in the sample data set to obtain non-redundant data; A partitioning module is used to divide the non-redundant data into training set data and test set data according to the proportion and time order; A training module is used to construct an AM-GRU-BPNN prediction model, train the AM-GRU-BPNN prediction model, and obtain a trained AM-GRU-BPNN prediction model; wherein the AM-GRU-BPNN prediction model includes a temporal attention mechanism AM, a gated recurrent unit network prediction model GRU, and a BP neural network prediction model; The second acquisition module uses the trained AM-GRU-BPNN prediction model to test the test set data and obtain the prediction results.

6. A terminal device, characterized in that: It includes a memory, a processor, and a program stored in the memory and executable by the processor. When the processor executes the program, it implements the product quality index prediction method based on AM-GRU-BPNN as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein when the program is executed, the device where the computer-readable storage medium is located is controlled to execute the product quality index prediction method based on AM-GRU-BPNN as described in any one of claims 1 to 4.

Images